{"id":"2248cc29-52eb-400f-ae0d-59e91eee22a3","arxiv_id":"2607.20554","paper_version":1,"verdict":"CONDITIONAL","confidence":"MODERATE","novelty_score":5.0,"correctness_risk":"medium","formal_verification":"none","parameter_count":5,"one_line_summary":"An edge-aware GNN trained on MILP labels selects multi-hop V2X relays with roughly 92% of the optimal connectivity gain at orders-of-magnitude lower runtime in simulation.","lead":"This paper trains a graph neural network to pick multi-hop relay routes for connected vehicles in cities, using an offline MILP solver's optimal decisions as training labels. It reports near-MILP connectivity at millisecond inference speeds in simulated Rome districts, arguing that real-time relay selection is feasible without exact optimization.","discovery_kind":"new_application","skeptic_critique":{"model":"deepseek-v4-flash","headline":"The MILP oracle is not actually a well-defined MILP: Eq. (7) forbids any active source-to-RSU path, and Eqs. (9)-(10) are non-linear; the reported 12.3% gain and all GNN labels rest on an unstated formulation.","rationale":"The reader's weakest assumption targets external validity (simulation fidelity), but the more immediate threat is internal validity: the published MILP formulation is internally inconsistent and incomplete. A supervised L2O claim depends first on the quality and correctness of the oracle labels; if the oracle formulation is not actually a MILP or admits only trivial solutions, then the reported oracle upper bound and the GNN's approximation ratio are ungrounded. This is not merely a missing artifact; it is a central technical gap in the argument as written. I still recommend the same conditional verdict the reader reached, but the conditions must include a corrected, executable MILP formulation and a reproducibility check on a small instance. My disagreement with the reader is about which assumption is most load-bearing, not about the overall verdict.","tokens_in":8622,"tokens_out":6686,"duration_ms":70245,"concrete_test":"Run the stated MILP (Eqs. 5-11) on a minimal snapshot with one RSU, one source CAV, and one relay CAV using Gurobi or SCIP, with f and x as variables. Because Eq. (7) forces inflow=outflow at the source with no injection, the solver should return objective 0; if it instead returns a connected path, inspect the actual constraint set and publish the corrected equations. Then re-derive Table II with that corrected formulation. If the oracle gain shifts materially from 12.3%, the GNN comparison and the 11.3% headline are not supported.","verdict_should_be":"UNCHANGED","load_bearing_attack":"The paper's central claim is that GINE recovers 11.3% of the MILP oracle's 12.3% connectivity gain, which requires the oracle to be a well-posed MILP whose optimal solutions are valid relay forests. The printed formulation in Section III-C is not that. Eq. (7) imposes flow conservation on every non-RSU node with no source injection or sink absorption, and f_ij is not among the decision variables listed in Eqs. (3)-(4). In any non-empty forest, a leaf CAV has zero incoming flow and positive outgoing flow, violating Eq. (7); the only feasible solution is the all-zero solution. Thus the formulation as written cannot produce the reported oracle gain. Additionally, Eq. (9) uses a path-existence indicator and Eq. (10) imposes acyclicity, neither of which is linear, so the problem is not a MILP as stated. Since the oracle's labels are the supervision signal for GIN/GINE, any gap between the printed constraints and the solver actually used means the headline near-oracle performance cannot be verified or reproduced from the manuscript. The reader's simulation-fidelity concern is valid but secondary: even with perfect SUMO-GEMV2 traces, the oracle labels are undefined without a corrected formulation.","agreement_with_reader":"disagree"},"referee_report":{"model":"deepseek-v4-flash","summary":"The paper proposes a Learning-to-Optimize (L2O) framework for multi-hop relay selection in dense urban NR-V2X networks. Vehicular snapshots are represented as attributed graphs, and a GINE/GIN model is trained in a supervised manner on labels generated by an offline MILP oracle. The authors report that the GINE model achieves an average connectivity gain of 11.3% versus the oracle's 12.3% over a single-hop baseline, while running in milliseconds, and that it outperforms a greedy SNR heuristic. Evaluation is based on 286,602 snapshots from a SUMO–GEMV2 simulation pipeline across three Rome districts, including cross-area generalization tests.","tokens_in":8967,"tokens_out":7949,"duration_ms":82564,"significance":"If the results hold, the paper would be a useful contribution: it demonstrates that an edge-aware GNN can approximate exact optimization decisions for a combinatorial relay-selection problem at a fraction of the solver's runtime, with a large-scale realistic simulation, explicit comparison against an oracle and a heuristic, and cross-area robustness tests. The held-out and cross-area evaluation design is a strength, and the authors are transparent that the evaluation is simulation-based. However, the entire empirical claim rests on the validity of the MILP oracle, and the printed formulation has serious correctness problems that currently prevent verification of every reported number.","major_comments":[{"comment":"The optimization problem as written is not a valid MILP and cannot generate the reported oracle labels. Variables f_ij used in (7)–(8) are not declared in (3)–(4). Constraint (7) imposes flow conservation on every non-RSU node without source injection or sink absorption; combined with the single-outgoing-link constraint (6), any nonempty forest has leaf CAVs with outgoing but no incoming flow, so (7) forces the all-zero solution. In addition, (9) uses a path-existence indicator and (10) imposes acyclicity as a global constraint; neither is linear. Since all GNN labels and the 12.3% oracle gain are produced by 'the MILP oracle' in Section IV-A, the discrepancy between the printed formulation and the actual solver makes the central claim unverifiable. Please provide the full formulation actually solved, including flow variables, demand injection, sink absorption, and a linearized connectiv","section":"§III-C, Eqs. (3)–(10)"},{"comment":"The runtime feasibility enforcement is not sufficient to guarantee the capacity and connectivity constraints of Section III. The text states that post-processing enforces single-outgoing-link selection and directed cycle elimination, but it does not mention capacity constraint (8) or demand feasibility. A thresholded GNN output may select a relay whose incoming demand exceeds the outgoing link capacity, producing an 'acyclic RSU-rooted forwarding graph' that violates the oracle's feasibility conditions. Please describe how capacity feasibility is enforced, or report the fraction of snapshots where the post-processed solution violates (8).","section":"§IV-C"},{"comment":"The reported gains are internally inconsistent as presented. Section V-C reports an average connectivity gain of 11.3% (GINE) over the whole dataset, while Section V-G reports median gains of 83.3% (suburban), 33.3% (urban-medium), and 19.8% (urban-stress), with improvements in 97% of snapshots. If the latter use the same metric (relative increase over single-hop V2I), an average of 11.3% is difficult to reconcile; if they are a different quantity, the metric needs to be defined. Please reconcile these numbers and provide the distribution or counts of snapshots per regime.","section":"§V-C and §V-G"},{"comment":"The MILP solver is not named; no version, optimality gap, time limit, or hardware are reported. The claimed oracle upper bound and the speedups in Table I depend on these details. Please specify them and report solver optimality gaps. This is especially important because the oracle's correctness is the basis for the supervised labels. Without these details, the reproducibility of the experiments is limited.","section":"§V-A, §V-D"}],"minor_comments":[{"comment":"'more than 9.5% of snapshots exceed 25.8% gain' should be '9.5% of snapshots exceed 25.8% gain' or similar; the current phrasing is ambiguous.","section":"§V-C"},{"comment":"H_max is optional and 'hop(i,j)' is not defined. Clarify whether this constraint was used in the experiments and, if so, how the hop count is computed.","section":"§III-C, Eq. (11)"},{"comment":"The composition of the 286,602 snapshots by regime (suburban, urban-medium, urban-stress) is not given. Providing counts would allow the reader to weight the Section V-G medians and reconcile them with the overall average.","section":"§V-A"},{"comment":"The threshold for converting GNN edge probabilities to active links is not specified. Report the threshold or describe how it is selected, since connectivity results depend on it.","section":"§IV-C"},{"comment":"The graph is directed, but the GIN/GINE message-passing description does not specify how edge direction is handled. Please clarify whether messages are aggregated over both directions or only along outgoing/incoming edges.","section":"§IV-B"},{"comment":"The 'Rel. Drop' for the 'All Areas' row (2.6%) is ambiguous; state the reference value used to compute the relative drop.","section":"Table II"},{"comment":"The greedy SNR baseline should specify how capacity and hop constraints are enforced and how ties are broken, so that the comparison is precisely reproducible.","section":"§IV-D"},{"comment":"The 'Worst-Case' scenario is not defined in Section V-A. Clarify what it represents and how it differs from 'Urban-Stress'.","section":"§V-D, Table I"}],"recommendation":"major_revision","confidential_remarks":"The invalid MILP formulation is severe: it affects the oracle labels, the upper bound, and every GNN result derived from those labels. I recommend a major revision only if the authors can supply the exact formulation actually solved and confirm that the reported numbers derive from it. If the corrected formulation does not reproduce the reported oracle gain, the paper should be rejected. It would also help to request code/data or a detailed reproducibility appendix, because the current text does not allow verification of the core experiments."},"author_rebuttal":null,"desk_editor":{"model":"deepseek-v4-flash","letter":"Dear colleague,\n\nQuick read of arXiv:2607.20554. The headline: don't trust the oracle numbers until the MILP is rewritten. Eq. (7) imposes equality of incoming and outgoing flow for every non-RSU node, with no source injection or sink absorption; the only feasible flow vector is all zero. So the printed optimization problem cannot produce the 12.3% gain reported in Section V-C. Eqs. (9) and (10) are not linear constraints either—path-existence and acyclicity are not MILP-representable as written. The authors clearly solved some other formulation; they just didn't write it down. That is the paper's load-bearing flaw: every GNN label and every comparison to the oracle inherits the unstated MILP.\n\nWhat's genuinely there: the study is competently executed engineering. The GINE idea—edge-aware message passing that folds SNR/LoS/distance into relay selection—is a natural but real increment over prior GNN-V2X work, and the cross-area evaluation (train on one Rome district, test on two others) is an honest robustness check. The runtime claims (millisecond inference vs MILP seconds) are plausible. If the MILP were properly specified, this would be a publishable within-subfield contribution.\n\nBeyond the MILP: no code or data, no error bars, no comparison to a non-graph ML baseline (e.g., gradient boosting on edge features), and the whole evaluation is simulation-only, which the authors admit in Section V-F. The domain-shift numbers in Table II are interesting, but they come from a single seed, so the relative drops are hard to interpret.\n\nBottom line: major revision. The paper deserves a serious referee because the problem is real and the approach is promising, but the central reference point is undefined as printed. If the authors provide the actual MILP (with flow-balance source terms and cycle-elimination cuts, plus a solver trace), the results become checkable. I wouldn't cite it in its current form.","headline":"The GNN-L2O framing is worth a look, but the printed MILP oracle is not a MILP: Eq. (7) forces zero flow, and the reported 12.3% oracle gain sits on an unspecified formulation.","tokens_in":9468,"tokens_out":2979,"would_cite":false,"duration_ms":30957,"reading_group":"yes","serious_thinker":"no","would_accept_peer_review":true},"pith_extraction":{"msc":[],"pacs":[],"model":"deepseek-v4-flash","headline":"A graph neural network approximates the optimal multi-hop relay choice for urban V2X at millisecond latency, recovering 11.3% of the ideal connectivity gain.","keywords":["NR-V2X","relay selection","graph neural networks","learning-to-optimize","MILP","GINE","urban vehicular networks","connectivity"],"falsifier":"Deploy the trained model in a real urban NR-V2X testbed with live RSUs and vehicles, record the actual connectivity and throughput, and recompute the MILP solution on the same snapshots. If the real-world connectivity gap between GINE and the MILP oracle exceeds the roughly one percentage point seen in simulation — or if the latency budget is missed — the central claim fails.","tokens_in":8539,"feed_emoji":"📡","tokens_out":2911,"duration_ms":27989,"temperature":0.7,"pith_summary":"The paper tries to show that a supervised graph neural network can learn the relay-selection decisions of an optimal but slow mixed-integer linear program in dense urban vehicle-to-everything networks. The network is trained on snapshots where the MILP labels the optimal active links, and at run time the GNN (specifically the edge-aware GINE variant) outputs a near-identical forwarding forest. On simulated Rome districts, this recovers most of the MILP's connectivity gain (11.3% vs 12.3% over single-hop) in a few milliseconds — orders of magnitude faster than solving the MILP. If correct, this makes optimal multi-hop relaying feasible in real time without new roadside hardware.","feed_headline":"GNN recovers 11.3% connectivity gain in milliseconds","feed_subtitle":"A learned graph policy approximates the optimal multi-hop relay solver on dense urban snapshots, cutting runtime by orders of magnitude.","key_machinery":"The key machinery is the edge-aware Graph Isomorphism Network (GINE) trained under a learning-to-optimize paradigm. Each vehicular snapshot becomes a directed attributed graph; GINE's message passing incorporates edge-level radio features, and an MLP head scores each candidate link. A linear feasibility refinement enforces single-outgoing-link and acyclicity constraints, turning scores into a valid RSU-rooted forest. The MILP solver is used only offline, to generate labels and an upper bound.","core_discovery":"The paper claims that relay selection in NR-V2X networks, which is a combinatorial optimization problem with flow, capacity, and acyclicity constraints, can be learned end-to-end from an MILP oracle. The central empirical discovery is that an edge-aware Graph Isomorphism Network (GINE), which injects radio characteristics (SNR, LoS/NLoS, distance, capacity) into the message-passing, matches the oracle's connectivity within one percentage point on average (11.3% vs 12.3% gain), while keeping inference at a constant few milliseconds. The topology-only GIN does slightly worse (10.78%), showing that radio-aware edge features carry useful signal.","pith_inferences":["If the GNN's decisions degrade gracefully under distribution shift, the same approach could be applied to dynamic resource allocation or routing in other wireless systems, not just vehicle-to-infrastructure.","A natural testable extension is to compare GINE's selected relays against a live MILP recomputation on streaming field data; the one-point gap observed in simulation is the quantity that must survive real RF conditions.","Because the model is supervised by an oracle, its performance ceiling is bounded by the MILP formulation's own simplifying assumptions (capacity based on Shannon, etc.). Better oracles would likely raise the ceiling."],"forward_implications":["Real-time multi-hop relay selection becomes feasible in dense urban scenarios without solving MILP online.","Operators can extend coverage to CAVs that currently have no single-hop path, using vehicles as relays rather than deploying more RSUs.","Edge-aware message passing gives a measurable edge over topology-only models, suggesting radio features should be embedded in GNN design for wireless problems.","Cross-area results indicate a model trained on multiple districts generalizes better; a single-district model loses 15–23% of its gain on unseen morphologies."],"fun_headline_variants":["GNN relay picks match MILP in milliseconds","Millisecond relay selection matches optimal MILP","Learned relay selection: near-optimal, orders faster","AI relay picker: 11.3% gain at millisecond speed","Graph nets cut V2X relay selection to milliseconds"],"cache_read_input_tokens":2304,"weakest_assumption_plain":"The results assume the simulated urban mobility and radio propagation pipeline faithfully reproduces real dense-city NR-V2X conditions; if real links differ, the trained labels and the claimed gain will not transfer.","fun_headline_variants_meta":{"raw":{"variants":["GNN relay picks match MILP in milliseconds","Millisecond relay selection matches optimal MILP","Learned relay selection: near-optimal, orders faster","AI relay picker: 11.3% gain at millisecond speed","Graph nets cut V2X relay selection to milliseconds"]},"model":"deepseek-v4-flash","effort":"low","cost_usd":0.000228,"raw_usage":{"total_tokens":1351,"prompt_tokens":821,"completion_tokens":530,"prompt_tokens_details":{"cached_tokens":256},"prompt_cache_hit_tokens":256,"prompt_cache_miss_tokens":565,"completion_tokens_details":{"reasoning_tokens":451}},"tokens_in":565,"tokens_out":530,"duration_ms":6142,"temperature":1.0,"reasoning_tokens":451,"cache_read_input_tokens":256,"cache_creation_input_tokens":0},"cache_creation_input_tokens":0},"created_at":"2026-08-02T02:20:40.047001+00:00","model_set":{"reader":"deepseek-v4-flash"},"falsifier":"Deploy the trained model in a real urban NR-V2X testbed with live RSUs and vehicles, record the actual connectivity and throughput, and recompute the MILP solution on the same snapshots. If the real-world connectivity gap between GINE and the MILP oracle exceeds the roughly one percentage point seen in simulation — or if the latency budget is missed — the central claim fails.","supporting_citations":[],"review_version":1}